tonic/xds: wire up lbv2 as a feature

[gu0keno0]

Add tonic-xds' own LB module as a feature

Motivation

Tower's LB is limited. We cannot support advanced xds features such as customized load balancer and outlier detection on top of it.

We have been developing tonic-xds' own LB for a while, and this PR make it ready for future consumptions.
We will still need to wire up xds configs for OD and RingHash, those can be done in future PRs.

Solution

Added the wire-up modules to make the tonic-xds' LB usable behind a feature flag.

[YutaoMa]

Found one potential issue with the new LB's readiness logic that needs verification:

1. LoadBalancer::poll_ready treats non-empty self.ready as sufficient readiness check, but when a cluster is dropped it doesn't evict from self.ready, leaving them stale forever. If those endpoints are down call will fail forever and even if new CDS config comes in using the same config name, because the code will reuse the existing registry entry, it would continue to reuse the broken set of self.ready instead of rebuild one.
2. The old tower::Balance path is less problematic because it re-polls readiness on every call and evict broken endpoints.

I think to fix it we either need to do eviction in call and/or rebuild entry in ClusterLbRegistry.

[gu0keno0]

Found one potential issue with the new LB's readiness logic that needs verification:

1. LoadBalancer::poll_ready treats non-empty self.ready as sufficient readiness check, but when a cluster is dropped it doesn't evict from self.ready, leaving them stale forever. If those endpoints are down call will fail forever and even if new CDS config comes in using the same config name, because the code will reuse the existing registry entry, it would continue to reuse the broken set of self.ready instead of rebuild one.
2. The old tower::Balance path is less problematic because it re-polls readiness on every call and evict broken endpoints.

I think to fix it we either need to do eviction in call and/or rebuild entry in ClusterLbRegistry.

poll_ready calls poll_discovery, so IMO it is poll_discovery's responsibility to handle cluster removal / adding back. Currently this is not working because cluster discovery is not re-established. I have a TODO here: https://github.com/gu0keno0/tonic/blob/39cda4e88d2e532581c3b9749c7150bb726c513f/tonic-xds/src/xds/lb_discovery.rs#L17-L19 .

This is also an issue for tower-lb as well: https://github.com/grpc/grpc-rust/blob/master/tonic-xds/src/xds/cluster_discovery.rs#L179 . If cluster_watch is closed (I think cluster removal triggers it today), cluster discovery dies.

So both tower-lb and lbv2 are under the same issue.